ChatGPT Work: What It Actually Does (and What It Replaced)
If you remember ChatGPT having an "agent mode" you could toggle on to have it click through websites and complete multi-step tasks on its own, and you've gone looking for that toggle recently and come up empty, you're not missing a setting. That broad, autonomous agent mode is no longer part of the current lineup. What replaced it isn't a single feature with the same shape. OpenAI points to ChatGPT Work for longer multi-step tasks and finished deliverables, and to cloud browser (a Work capability, available on paid plans in supported regions, not on Free or Go) for browser workflows.
What actually happened
Agent mode's original pitch was broad: describe a task, and ChatGPT would go do it, browsing sites, filling forms, clicking through a checkout flow, more or less unsupervised. In practice, that kind of open-ended autonomous browsing turned out to be the harder and riskier half of the job, the part most likely to go wrong in a way you wouldn't catch until it already had. OpenAI hasn't spelled out every reason, but the practical result is that the concept now lives in two narrower tools instead of one broad one.
If a tutorial mentions turning on 'agent mode'
That instruction is outdated now. There's no agent mode toggle to find. The functionality lives in ChatGPT Work and its cloud browser instead, and they work differently enough that it's worth understanding both before assuming either one does what agent mode used to.
Laid out side by side, the split looks like this:
| Agent mode (retired) | ChatGPT Work | Cloud browser | |
|---|---|---|---|
| Scope | One broad tool covering both research and live-site actions | Multi-step research, drafting, and compiling into one output | Navigating and acting on a specific live website |
| Autonomy | High, largely unsupervised end to end | Structured; still meant to be reviewed before you act on the result | Structured and scoped to the site or flow you point it at |
| Typical output | Inconsistent: sometimes a completed action, sometimes one that stalled partway | A document, table, or draft you read and use | A completed on-site action, like a submitted form or a checked status |
| Best for | (Retired for a reason: this breadth is what made it unpredictable) | Self-contained work that doesn't need to touch a live site | Work where the point is the interaction with the site itself |
| Status | Retired | Current | Current, on paid plans in supported regions |
What ChatGPT Work actually covers
ChatGPT Work is built for multi-step tasks: the kind of request that has several distinct parts done in sequence, where the value is in ChatGPT handling the whole sequence rather than you copy-pasting between separate chat requests. Think of it as the tool for work that's structured and mostly self-contained, research a topic, draft a document, compile a set of files, produce a report, rather than work that requires reaching out into the live, changing web and interacting with specific sites.
A realistic example: "Research three competitor pricing pages, summarize the differences in a table, and draft an email to the team recommending a pricing change" is the shape of task ChatGPT Work is built for. Each step feeds the next, and the whole sequence produces one coherent deliverable at the end.
What moved to the cloud browser feature instead
The cloud browser is the narrower descendant of what agent mode used to do with live websites: navigating an actual site, clicking through a specific flow, interacting with a page the way a person would. It works as part of ChatGPT Work and is available on paid plans in supported regions, not on Free or Go. Use it specifically when the task genuinely requires operating inside a real website rather than gathering and synthesizing information about it. Work can also read public pages for research on its own, so you don't need the cloud browser just to look things up. What each plan gets, and how much usage, changes over time, so check your plan's current limits.
The distinction that actually matters day to day:
ChatGPT Work
- Multi-step tasks that produce a document, report, or written output
- Research, drafting, and compiling done in sequence
- Best when the task is self-contained and doesn't require live site interaction
Cloud browser
- Tasks that require actually navigating a live website
- Clicking through a specific flow, form, or interface
- Best when the value is in the interaction itself, not just information gathered from it
How to tell which one a task actually needs
Ask yourself what the finished output looks like. If it's a document, a summary, a draft, a compiled report, something you'd read or send, that's ChatGPT Work's territory even if part of the task involves looking things up online along the way. If the point of the task is the action itself happening on a specific site, submitting a form, completing a multi-step signup, checking a status that only exists inside a logged-in dashboard, that's squarely cloud browser territory.
Plenty of real tasks are actually a combination: research done through ChatGPT Work, followed by a specific action that needs the cloud browser. Splitting a task like that into its two natural pieces usually gets a more reliable result than expecting one tool to do both smoothly, especially now that the old, broader agent mode is gone.
Be honest about what's actually reliable right now
Both of these are still tools for tasks where you review the outcome rather than tools you should treat as fully unsupervised. For anything involving a payment, a legal commitment, or an irreversible action, keep a human review step before anything gets submitted for real, regardless of which tool did the groundwork.
A realistic use case, worked through
Say you're a small operations team lead who needs to gather vendor quotes for a software renewal, summarize them, and flag which one needs a follow-up question before your Friday budget meeting.
I need to compare three vendor quotes for renewing our project management software: [paste or describe the three quotes]. Summarize the pricing, contract length, and any notable terms in a table, flag anything that looks like it changed unfavorably from our current contract, and draft two follow-up questions I should send to whichever vendor has the least clear terms.
”Here's roughly what a finished ChatGPT Work run on that prompt actually hands back, all three steps done in sequence and delivered as one output:
ChatGPT
ChatGPT Work's finished output, illustrated
I've compared the three renewal quotes below.
| Vendor | Annual price | Contract length | Notable term |
|---|---|---|---|
| Northstar PM | Up 12% vs. current | 2 years | Price locked for the full term |
| Fieldway Suite | Up 4% vs. current | 1 year | Auto-renews unless cancelled 60 days out |
| Corridor Tools | Up 18% vs. current | 1 year | "Implementation support" listed but scope undefined |
Corridor Tools has the least clear terms: the price increase is the steepest, and "implementation support" isn't defined anywhere in the quote. Here are two follow-up questions worth sending them before Friday:
First, what specifically is included under "implementation support," and is it a one-time service or ongoing? Second, is the 18% increase tied to new features we're not currently using, or is it a straight renewal increase on the same plan?
That's a ChatGPT Work-shaped task: multiple steps, one coherent output, no need to actually log into anything live. If the next step were "now go check the vendor's client portal to see if our renewal date is listed correctly," that specific piece would be cloud browser territory instead, since it requires navigating an actual logged-in site, assuming your plan and region support it and the site is one it can work with.
What this means if you were relying on the old agent mode
If your workflow depended on ChatGPT autonomously completing an end-to-end task that spanned research and live-site actions in one unsupervised pass, that specific broad capability is genuinely gone, not renamed. The honest framing is that one broad tool became two narrower ones, and the narrower ones don't necessarily cover everything the old one did, so test your own workflow before relying on it. That's a real change in what's possible in one step, not just a rebrand, and it's worth adjusting your expectations rather than assuming the old behavior is hiding somewhere under a new name.
If you're getting oriented on ChatGPT's current feature set more broadly, the Complete Beginner's Guide to ChatGPT is a solid starting point. And if this is one of a few 2026 product changes you're trying to track down, you're not alone: ChatGPT's voice mode went through a similar reset around the same time (it's now GPT-Live), and the old side-panel drafting tool became Writing blocks. All three changes landed within a few months of each other, which is part of why the confusion about what's current has been so widespread.
Official sources
Checked on September 21, 2026. Features, plans and names change often, so the vendor's own pages are the final word.